AI/Infra: Add memory caching layer for commonly evaluated positions

[chinweobtagaz]

Description

Many users evaluate the exact same positions (e.g., standard opening deviations or popular puzzles). Re-evaluating them is a waste of resources.

What Needs to be Done

• Implement a Redis cache layer for the engine worker orchestrator.
• Hash the FEN and requested depth as the cache key.
• Store the best move, evaluation, and principal variation (PV).

Acceptance Criteria

• Cache hits return instantly without querying the engine pool.
• Cache automatically evicts old/unused positions (LRU policy).
• Handles millions of keys efficiently with minimal memory footprint.